Search Systems
The half-second between keystroke and results is where sales are won. One 350M specialist and a token-level retriever run the whole pipeline at the edge, with no cloud API in the hot path.
The catalog behind every stage
Cedar Supply Co. is a fictional store with 100 products across 16 categories. Expand, Retrieve, Rescue, and Re-rank all search these same shelves — the counts and product names you see are real, so check them yourself. Understand is the one stage that never touches the index: it runs first, on the query alone.
The models never hold the catalog — they read the query. Products live in the index. That is why going from 100 items to a marketplace with millions is a re-indexing job, not a retraining one: the same specialist runs unchanged, and only the index grows.
talk trackNothing has hit the index yet. A shopper ready to buy, a shopper still comparing, and someone hunting for store hours should not go to the same place — this classifier decides which one just arrived.